The short answer
Yes. Every large language model hallucinates. The rate depends on the model, the task, and most importantly, the benchmark you are measuring against. Anyone who tells you a specific model "does not hallucinate" is wrong, including the marketing team that may have told them that.
What hallucination is, technically
Large language models do not retrieve facts. They predict the next most-likely token given a sequence of prior tokens. When the training corpus has covered a topic densely and consistently, the predicted tokens line up with what we recognize as the truth. When the training corpus is sparse, contradictory, or absent, the prediction is plausible-sounding language that is not grounded in fact. That is a hallucination. It is not a malfunction. It is the architecture doing exactly what it was trained to do.
The Vectara leaderboard, 2026
The Vectara hallucination leaderboard measures summarization hallucination -- give the model a passage, ask for a summary, count the unsupported claims. As of May 2026:
- Gemini 2.0 Flash: 0.7%
- Gemini 3 Pro: about 1.0%
- GPT-5: 1.4%
- Claude Sonnet 4.5: 4.4%
- Claude Opus 4.6: about 4.9%
- Llama 4 Maverick: 4.6%
- DeepSeek R1: 11 to 14%
- Grok 4.1 Fast: about 72% (specific summarization tasks)
These numbers move week to week. Check the leaderboard for current values.
The AA-Omniscience benchmark and why it matters more
Vectara measures summarization, where the source is in the prompt and the model just has to be faithful to it. AA-Omniscience measures something harder: open-knowledge questions, where the model has nothing in front of it and must know whether it knows. On that benchmark, GPT-5.5 was 57% accurate on questions it could answer, and 86% likely to confidently fabricate when it could not. The 86% omniscience rate is the number to worry about for any task where the model is being asked to recall from its training, which is most real-world use.
RISK
If you ask an LLM "what cases support my position" without giving it the cases as input, the model is operating in the AA-Omniscience regime, not the Vectara regime. The relevant hallucination rate could exceed 80 percent. Verify before you cite.
The Mata v Avianca lawyer case
In 2023, an attorney for the plaintiff in Mata v Avianca submitted a federal court brief citing six precedent cases. None of the cases existed. ChatGPT had fabricated all six, and the attorney had not verified them against any primary source. The court sanctioned him, fined him, and the case is now part of the legal ethics curriculum. It is the canonical example of what happens when a professional treats LLM output as a primary source. The lesson generalizes: any high-stakes use of AI output requires a verification step against a primary source. There are no shortcuts.
How Restrepo guards against hallucination
We do four things on every client deployment:
- Verification step. AI output that will be transmitted to a third party is reviewed by a human, every time.
- No auto-send on irreversible actions. The system can draft an email; the human clicks send. The system can draft a calendar invite; the human accepts. The system can recommend a reservation; the human confirms.
- Manual override on safety-relevant control. No LLM autonomously controls doors, alarms, water valves, gas, or moves money.
- Citation at the source level. When we route through Perplexity for research, we click the citation. When we route through Claude for drafting, we verify any factual claim before publication.
The Perplexity 37% finding
Even citation-first answer engines hallucinate. A Columbia Journalism Review audit found 37% of Perplexity answers contained errors despite the citations. The citations show the model looked at sources. They do not show the model summarized them correctly. We use Perplexity daily. We say this anyway because it is true. (See our Perplexity review.)
Why prompts cannot fix it
The most common advice you will read on the internet is that hallucinations can be reduced with better prompting. "Tell the model to say I do not know." "Use chain-of-thought." "Add a system prompt that says do not invent." These help, marginally, on the margin. They do not solve the problem because the problem is architectural. The model produces fluent next tokens; whether the tokens are true depends on the training data. No prompt can bring back a fact the model never saw.
What does help, structurally, is retrieval-augmented generation: give the model the source material as input and tell it to summarize that material rather than recall its own training. This is what Perplexity does. This is what we do internally for client work. It cuts the hallucination rate by an order of magnitude on factual questions. It does not cut it to zero, because the model can still misinterpret what it is shown. The defense is verification, not prompting.
What this means for you
It does not mean do not use AI. It means do not use AI as the last word. Use it as a first draft, a research head start, a transcription engine, a scene generator. Verify anything that has consequences. Build the verification into the workflow so it is automatic rather than aspirational.